In low-light scenarios such as nighttime inspection and post-disaster search and rescue, robotic-arm teleoperation is often hindered by insufficient illumination, which leads to missing geometric details and degraded textures in the scene. This severely limits the accuracy and reliability of 3D reconstruction and, in turn, compromises the execution of teleoperation tasks. Existing low-light enhancement methods mostly focus on improving the perceptual quality of single-view images, while lacking explicit modeling of intrinsic multi-view consistency. As a result, downstream 3D reconstruction suffers from bottlenecks such as failed feature matching and cross-view artifacts. To address these challenges, this paper proposes a unified reconstruction framework that integrates multi-view consistent diffusion enhancement with 3D Gaussian Splatting (3DGS). Specifically, a joint optimization mechanism is introduced into the diffusion model by incorporating 2D reconstruction regularization and a cross-view consistency loss. This design enhances image illumination and recovers scene details while strictly enforcing high consistency in photometric features across views. Furthermore, the enhanced multi-view image sequences are leveraged to achieve efficient and high-fidelity scene-level 3D Gaussian geometric representation. Finally, the reconstructed 3D scene representation is seamlessly integrated into a virtual reality platform to support immersive interaction and teleoperation in complex environments. Experimental results demonstrate that, in real-world low-light scenes, the proposed method not only significantly improves image visual quality and 3D reconstruction accuracy, but also provides precise visual perception for robotic-arm teleoperation and grasping tasks, thereby validating its strong application potential in low-light environments.

Lifting 2D segmentation masks into 3D Gaussian Splatting (3DGS) representations is a promising direction for efficient scene understanding. While existing lifting approaches have demonstrated potential in real-time applications, they face non-trivial challenges regarding semantic ambiguity and geometric boundary precision. Specifically, relying solely on gradient accumulation often leads to suboptimal distinction between targets and semantic distractors, as well as spatial leakage along object boundaries due to the lack of explicit geometric constraints. To address these challenges, we propose ConSplat, a novel training-free framework that enforces a Dual-Consistency Strategy for robust open-vocabulary segmentation. First, we introduce a Gradient Consistency (GC) module. Distinct from traditional binary voting, this module explicitly models and suppresses semantic distractors, ensuring that the lifting process is guided by exclusive target identity. Second, to resolve boundary ambiguity, we incorporate a 3D-2D Projection Consistency (PC) metric. This mechanism imposes a strict geometric constraint by measuring the alignment between projected Gaussians and 2D masks, effectively pruning primitives that lack spatial fidelity. Finally, we fuse these semantic and geometric cues via Density Peak Clustering to automatically identify target primitives without manual thresholding. Extensive experiments on LERF-Mask and 3D-OVS datasets demonstrate that ConSplat achieves superior segmentation accuracy. Remarkably, our method operates approximately 13 × faster than optimization-based baselines and maintains robust performance even under sparse-view settings.

Plant phenotyping is essential for modern crop breeding, yet traditional static image analysis fails to capture the nonlinear dynamics of plant growth. Existing time-series forecasting models exhibit notable limitations when processing multimodal data: global pooling operations may compress local 2D spatial topology of plants, and shallow feature concatenation may be insufficient for effective cross-modal semantic alignment. Moreover, current methods typically regress absolute morphological states, which may contribute to temporal lag during nonlinear growth spurts. In this paper, we propose ST-CrossGro-Former, a cross-modal residual forecasting framework for plant dynamic growth. The network removes the final global pooling and classification layers to preserve spatial topology and incorporates a scalar-guided cross-modal attention module based on the standard query-key-value formulation. This module utilizes 1D morphological features as queries to dynamically weight local visual regions, promoting multimodal feature alignment. Concurrently, a residual incremental forecasting strategy is introduced to predict short-term growth increments rather than absolute states, aiming to improve tracking sensitivity to sudden growth events. Evaluations on the UNL-CPPD maize dataset and supplementary validation on the FIP1 wheat field dataset show that the proposed model achieves competitive single-step forecasting accuracy and favorable temporal trajectory alignment compared with adapted spatiotemporal attention, graph-based, and physics-informed baselines under the evaluated settings. In particular, the FIP1 results suggest that ST-CrossGro-Former can maintain favorable height trajectory alignment under a field-acquired wheat setting, indicating its potential for helping mitigate temporal misalignment in dynamic growth forecasting.

Local navigation controllers can fail to make progress in constrained indoor environments when narrow geometry and moving obstacles compress the feasible local motion space. These freeze events trigger repeated recovery behaviors and can prevent otherwise valid global paths from being executed. This paper treats freezing as a failure of local control under three observable mechanisms: insufficient path horizon, motion ambiguity in single-frame sensing, and unstable learned velocity commands. We present MSTR-PPO (Multi-Scale Temporal Residual Proximal Policy Optimization), a freeze-aware residual local controller defined at the standard sensor--path--velocity-command interface. Rather than replacing the geometric controller with an unconstrained end-to-end policy, MSTR-PPO learns a bounded correction around a Pure Pursuit nominal action. The controller uses a Multi-Scale Path Transformer for near-, mid-, and far-horizon route context, a First-Order Temporal High-Pass Filter for low-cost light detection and ranging (LiDAR) range-change cues, and a Bounded Residual Action around Pure Pursuit. We evaluate eight local-control variants across nine Gazebo worlds with dynamic obstacles, totaling 2,880 trials across all methods, and deploy the trained policy through a Robot Operating System 2 (ROS, 2) Navigation 2 (Nav2) C++ controller interface as a representative evaluation stack. In the two target freeze-prone worlds, MSTR-PPO reaches the goal in all 80 evaluated trials and triggers no stack-level recovery events, while the best default-parameter classical baseline reaches 86.2% success on the same subset. On the main seven-world evaluation suite (W1--W7), MSTR-PPO obtains 91.1% average success and runs with 5--10, ms central processing unit (CPU) control-cycle latency. Stress tests expose a limitation on continuous-curvature paths, and a preliminary Unitree Go2 test succeeds in 27 of 30 trials.

Despite various methods in robotic operation, it remains a great challenge for industrial mobile phone flexible printed circuit (FPC) assembly due to the extremely stringent demand in position identification for the very tiny FPC assembly targets under strict assembly tolerance. To this end, this work proposes an accurate FPC position identification strategy in industrial robotic mobile phone assembly. The contributions of the work are concluded as follows: (1) we construct a Multi-Head Attention (MHA) architecture with the objective of encoding the visual-haptic information of the position into compact fusion presentations; (2) the distance and rotation errors between the ideal FPC position and the varying mismatched FPC positions around the correct one are connected to the Multi-Head Attention decoder information in a regression manner; (3) a dynamic weight averaging (DWA) strategy is adopted to adjust the weights in the loss calculations, which is able to achieve a better balance between the position and rotation errors in loss regression. Experiments were conducted on a practical FPC assembly platform. The results show that the proposed method can significantly improve the FPC location accuracy under the limited requirements of assembly attempts. The possibility exists that the proposed method applies to the real mobile phone assembly lines to reduce the labor burden in the near future.

World models (WMs) have demonstrated strong potential for end-to-end autonomous driving by learning predictive representations of future scene dynamics. However, generating future videos during inference introduces substantial computational overhead, leading many recent driving WMs to adopt a single front camera as input for efficient deployment. This design restricts spatial coverage in safety-critical maneuvers such as lane changes, merges, and turns. To address this limitation, we propose SV-WAM, a surround-view world-action model (WAM) that preserves full six-camera observations while maintaining efficient inference. SV-WAM leverages future-video prediction as dense training supervision for action learning within a shared generative model, rather than as an inference-time output. At the core of this design is an action-centered causal mask that prevents action tokens from attending to future-video tokens during joint action-video denoising. Consequently, the video branch can be discarded at deployment, enabling efficient action-only planning. Furthermore, we introduce a differentiable drivable-area compliance regularizer that penalizes vehicle-footprint corners approaching or crossing drivable boundaries, improving planning safety and boundary awareness. Extensive experiments on the closed-loop NAVSIMv2 benchmark and the open-loop nuScenes benchmark demonstrate that SV-WAM achieves state-of-the-art planning performance with low inference latency and competitive zero-shot transfer capability.

Enabling robots to learn manipulation tasks is a practical engineering application, and reinforcement learning is one of the key artificial intelligence methods to achieve it, but reinforcement learning always faces the trade-off between the ease of designing reward functions and the ease of learning from rewards. Reward shaping provides a solution and recent works have designed rewards using images and language descriptions, which is a simple and convenient shaping method for non-expert users. However, some of them only adopt the pretrained model without fine-tuning, while others train the reward model with absolute score labels, which all have difficulties in capturing the spatial relationships within the images, so the performance of reward shaping models is limited. In this work, we propose a novel reward shaping method to generate additional rewards from task descriptions and scene images. We utilize the pretrained vision-language model as backbone for efficient cross-modal information fusion and design a downstream task trained with pair-wise comparison for reward shaping. Extensive experiments are conducted to demonstrate the effectiveness of each component in the method. The approach is validated in the Meta-World environment and the results demonstrate that it outperforms standard reinforcement learning and the existing work in terms of the policy learning efficiency.

Deep reinforcement learning methods have shown promising results in learning specific tasks, but struggle to cope with the challenges of long horizon manipulation tasks. As task complexity increases, the large state space and sparse reward make it difficult to collect effective samples through random exploration. Hierarchical reinforcement learning decomposes complex tasks into subtasks, which can reduce the difficulty of skill learning, but still suffers from limitations such as inefficient training and poor transferability. Recently, large language models (LLMs) have demonstrated the ability to encode vast amounts of knowledge about the world and to excel in context-based learning and reasoning tasks. However, applying LLMs to real-world tasks remains challenging due to their lack of grounding in specific task contexts. In this paper, we leverage the planning capabilities of LLMs alongside reinforcement learning (RL) to facilitate learning from the environment. The proposed approach yields a hierarchical agent that combines LLMs with parameterized action primitives (LARAP) to address long-horizon manipulation tasks. Rather than relying solely on LLMs, the agent uses them to guide a high-level policy, improving sample efficiency during training. Experimental results show that LARAP significantly outperforms baseline methods across various simulated manipulation tasks. The source code is available at: https://github.com/ningzhang-buaa/LARAP-code.

Artificial intelligence (AI) and robots offer vast opportunities in shifting toward precision agriculture to enhance crop yields, reduce costs, and promote sustainable practices. However, many crop traits obstruct the application of AI-based robots. One bottleneck is flower morphology with recessed stigmas, which hinders emasculation and pollination during hybrid breeding. We developed a crop-robot co-design strategy in tomatoes by combining genome editing with artificial-intelligence-based robots (GEAIR). We generated male-sterile lines bearing flowers with exserted stigmas, and then trained a mobile robot to automatically recognize and cross-pollinate those stigmas. GEAIR enables automated F1 hybrid breeding with efficiency comparable to manual pollination and facilitates the rapid breeding of stress-resilient and flavorful tomatoes when combined with de novo domestication under speed-breeding conditions. Multiplex gene editing in soybean recapitulated the male-sterile, exserted-stigma phenotype, potentially unlocking robotized hybrid breeding. We demonstrate the potential of GEAIR in boosting efficiency and lowering costs through automated, faster breeding of climate-resilient crops.

Plant 3D reconstruction plays a critical role in precision agriculture and plant growth monitoring, yet it faces challenges such as complex background interference, difficulties in capturing intricate plant structures, and a slow reconstruction speed. In this study, we propose PlantSamGaussianReconstruction (PSGR), a novel method that integrates Grounding SAM with 3D Gaussian Splatting (3DGS) techniques. PSGR employs Grounding DINO and SAM for accurate plant–background segmentation, utilizes algorithms such as Scale-Invariant Feature Transform (SIFT) for camera pose estimation and sparse point cloud generation, and leverages 3DGS for plant reconstruction. Furthermore, a 3D–2D projection-guided optimization strategy is introduced to enhance segmentation precision. The experimental results of various multi-view plant image datasets demonstrate that PSGR effectively removes background noise under diverse environments, accurately captures plant details, and achieves peak signal-to-noise ratio (PSNR) values exceeding 30 in most scenarios, outperforming the original 3DGS approach. Moreover, PSGR reduces training time by up to 26.9%, significantly improving reconstruction efficiency. These results suggest that PSGR is an efficient, scalable, and high-precision solution for plant modeling.

The automation of assembly operations with industrial robots is pivotal in modern manufacturing, particularly for multispecies, low-volume, and customized production. Traditional programing methods are time-consuming and lack adaptability to complex, variable environments. Reinforcement learning-based assembly tasks have shown success in simulation environments, but face challenges like the simulation-to-reality gap and safety concerns when transferred to real-world applications. This article addresses these challenges by proposing a low-cost, image-segmentation-driven deep reinforcement learning strategy tailored for insertion tasks, such as the assembly of peg-in-hole components in satellite manufacturing, which involve extensive contact interactions. Our approach integrates visual and forces feedback into a prior dueling deep Q-network for insertion skill learning, enabling precise alignment of components. To bridge the simulation-to-reality gap, we transform the raw image input space into a canonical space based on image segmentation. Specifically, we employ a segmentation model based on U-net, pretrained in simulation and fine-tuned with real-world data, significantly reducing the need for labor-intensive real image segment labels. To handle the frequent contact inherent in peg-in-hole tasks, we integrated safety protections and impedance control into the training process, providing active compliance and reducing the risk of assembly failures. Our approach was evaluated in both simulated and real robotic environments, demonstrating robust performance in handling camera position errors and varying ambient light intensities and different lighting colors. Finally, the algorithm was validated in a real satellite assembly scenario, achieving a success rate of 15 out of 20 tests.

Today, skill transfer from humans to agents via demonstrations is a widely adopted approach. However, prior research primarily focused on recording data generated during demonstrations, leading to challenges such as poor interpretability of the skill transfer process, limited transferability and generalization capabilities of demonstrations, and insufficient visualization. To address these issues, this paper proposes a method for agent skill learning and knowledge graph generation based on demonstrations. The contributions include the following. First, to enhance the interpretability of the skill transfer process, we construct demonstration skills. Second, to improve the transferability and generalization of demonstrations, we propose a DeepSeek-based method for decomposing demonstration skills and generating behavior trees. Finally, to strengthen visualization, we introduce a hierarchical skill generation method for agents using knowledge graphs. Experiments conducted in 3C assembly and satellite assembly scenarios demonstrate that our method substantially enhances the interpretability of the demonstration process.

Deep reinforcement learning (DRL) algorithms have been widely applied to robotic manipulation tasks. However, state-of-the-art DRL methods struggle with complex multi-step tasks and sparse-reward settings due to inefficient exploration and reliance on predefined expert knowledge for sub-goal selection. To address these challenges, we propose a hierarchical reinforcement learning framework that integrates demonstrations to improve sub-goal discovery and exploration efficiency. Specifically, we introduce an object-centered sub-goal generation method that autonomously decomposes tasks into meaningful sub-goals by leveraging demonstration data. Furthermore, we enhance both high- and low-level policy learning through demonstration-guided DRL algorithms, improving sample efficiency and task generalization. We evaluate our approach in an open-source multi-step robotic manipulation environment. Experimental results demonstrate that our framework significantly outperforms state-of-the-art imitation learning and hierarchical reinforcement learning baselines in task completion success rates and learning stability.

Grasping unknown objects without re-training is still a challenging task for robot grasping. The traditional methods, including data-driven and transfer-learning based grasp methods still suffered from the extra data labeling, training on new samples. To this end, this paper proposes a method to quickly predict the grasping position of unknown objects. This method uses a Deep Siamese Network (DSN) as the backbone to estimate the similarity between the query object and the support set objects, and designs a new loss function through the shape feature estimation mechanism to learn the correct grasping position of unknown objects. Our method is compared to the state-of-the-art (SOTA) method on a well-known open grasp dataset: Connell Grasp Dataset, and a practical dataset. Experimental results show that the proposed method significantly surpasses the SOTA methods in the one-shot grasping task and has good generalization ability for unknown objects. Given only one unknown object sample, the proposed method can obtain a reliable grasping position of an unknown object without the need to re-collect data for retraining.

When presented with one or a few photos of a previously unseen object, humans can instantly recognize it in different scenes. Although the human brain mechanism behind this phenomenon is still not fully understood, this work introduces a novel technical realization of this task. It consists of two phases: (1) generating a Similarity Density Map (SDM) by convolving the scene image with the given object image patch(es) so that the highlight areas in the SDM indicate the possible locations; (2) obtaining the object occupied areas in the scene through a Region Alignment Network (RAN). The RAN is constructed on a backbone of Deep Siamese Network (DSN), and different from the traditional DSNs, it aims to obtain the object accurate regions by regressing the location and area differences between the ground truths and the predicted ones indicated by the highlight areas in SDM. By pre-learning from labels annotated in traditional datasets, the SDM-RAN can detect previously unknown objects without fine-tuning. Experiments were conducted on the MS COCO, PASCAL VOC datasets. The results indicate that the proposed method outperforms state-of-the-art methods on the same task.

Accurate detection of tomato fruits is a critical component in vision-guided robotic harvesting systems, which play an increasingly important role in automated agriculture. However, this task is challenged by variable lighting conditions and background clutter in natural environments. In addition, the arbitrary orientations of fruits reduce the effectiveness of traditional horizontal bounding boxes. To address these challenges, we propose a novel object detection framework named SN-YOLO. First, we introduce the StarNet’ backbone to enhance the extraction of fine-grained features, thereby improving the detection performance in cluttered backgrounds. Second, we design a Color-Prior Spatial-Channel Attention (CPSCA) module that incorporates red-channel priors to strengthen the model’s focus on salient fruit regions. Third, we implement a multi-level attention fusion strategy to promote effective feature integration across different layers, enhancing background suppression and object discrimination. Furthermore, oriented bounding boxes improve localization precision by better aligning with the actual fruit shapes and poses. Experiments conducted on a custom tomato dataset demonstrate that SN-YOLO outperforms the baseline YOLOv8 OBB, achieving a 1.0% improvement in precision and a 0.8% increase in mAP@0.5. These results confirm the robustness and accuracy of the proposed method under complex field conditions. Overall, SN-YOLO provides a practical and efficient solution for fruit detection in automated harvesting systems, contributing to the deployment of computer vision techniques in smart agriculture.

Recently, 3D Gaussian Splatting has gained significant attention in the novel view synthesis domain due to its high-quality rendering and computational efficiency. However, when camera views (e.g., focal length, scene depth) vary, the rendered images exhibit aliasing artifacts, including erosion during zoom-in and dilation during zoom-out. The fundamental cause of this phenomenon lies in the fixed-scale 2D dilation filter, which can not adapt to dynamic changes in sampling frequency. To address this issue, a mechanism for adaptive frequency constraint on Gaussians has been proposed, which dynamically adjusts frequency characteristics of Gaussians to match varying sampling requirements. Additionally, based on spectral discrepancies in images, a frequency regularization term is designed to enhance consistency in frequency domain representation. These approaches effectively suppress aliasing artifacts while maintaining rendering efficiency.

Multi-peg-hole assembly is a fundamental process in robotic manufacturing, particularly for circular aviation electrical connectors (CAECs) that require precise axial alignment. However, CAEC assembly poses significant challenges due to small apertures, posture disturbances, and the need for high error tolerance. This paper proposes a dual-stream Siamese network (DSSN) framework that fuses visual and tactile modalities to achieve accurate position identification in six-degree-of-freedom robotic connector assembly tasks. The DSSN employs ConvNeXt for visual feature extraction and SE-ResNet-50 with integrated attention mechanisms for tactile feature extraction, while a gated attention module adaptively fuses multimodal features. A bidirectional long short-term memory (Bi-LSTM) recurrent neural network is introduced to jointly model spatiotemporal deviations in position and orientation. Compared with state-of-the-art methods, the proposed DSSN achieves improvements of approximately 7.4%, 5.7%, and 5.4% in assembly success rates after 1, 5, and 10 buckling iterations, respectively. Experimental results validate that the integration of multimodal adaptive fusion and sequential spatiotemporal learning enables robust and precise robotic connectors assembly under high-tolerance conditions.

Various methods have been discussed in the field of Low-light Image Enhancement (LIE), however the inability to effectively restrain noise still led to unsupervised LIE remaining a challenge. Traditional supervised LIE methods had the advantage of suppressing noise, but they suffer from generalization problems since the pairing of low-light and normal-light images are always required in the process. To this end, this letter introduces an unsupervised LIE enhancement method that eliminates the need for paired images and effectively suppresses noise. The method utilizes the Retinex image decomposition technique to split the image into illumination and reflection components and processes the features of each part separately. In addition, we designed two deep learning architectures, lighting net (LINet) and denoising net (DNNet), to brighten the illumination component and denoise the reflection component respectively. We validate our method on two well-known open datasets and compared it with the state-of-the-art approaches. The experiments show that that the proposed method performs well in dimly lit scenes and visual verification, all quantitative metrics have at least 8% improvement on the best performing validation set.

Autonomous pollination robots have been widely discussed in recent years. However, the accurate estimation of flower poses in complex agricultural environments remains a challenge. To this end, this work proposes the implementation of a transformer-based architecture to learn the translational and rotational errors between the pollination robot’s end effector and the target object with the aim of enhancing robotic pollination efficiency in cross-breeding tasks. The contributions are as follows: (1) We have developed a transformer architecture model, equipped with two feedforward neural networks that directly regress the translational and rotational errors between the robot’s end effector and the pollination target. (2) Additionally, we have designed a regression loss function that is guided by the translational and rotational errors between the robot’s end effector and the pollination targets. This enables the robot arm to rapidly and accurately identify the pollination target from the current position. (3) Furthermore, we have designed a strategy to readily acquire a substantial number of training samples from eye-in-hand observation, which can be utilized as inputs for the model. Meanwhile, the translational and rotational errors identified in the end-manipulator Cartesian coordinate system are designated as loss targets simultaneously. This helps to optimize the training of the model. We conducted experiments on a realistic robotic pollination system. The results demonstrate that the proposed method outperforms the state-of-the-art method, in terms of both accuracy and efficiency.